Meta unveils four custom AI chips

Meta’s news here is not “we made a chip.” It’s “we’re turning custom AI silicon into a fast-moving product line.” On March 11, Meta said it is developing and deploying four new generations of its MTIA chips within the next two years, aimed at ranking, recommendation, and generative AI inference. That is a much faster cadence than normal chip cycles, and it tells you what Meta thinks the real bottleneck is now — not just training giant models, but serving them cheaply and fast at massive scale. ### What exactly did Meta announce? Meta laid out a roadmap for four upcoming generations of its Meta Training and Inference Accelerator family — often shortened to MTIA — with deployment stretching through the end of 2027. The company framed MTIA as central to its AI infrastructure, not a side experiment, and said these chips are being built specifically to support the workloads that power feeds, ads, recommendations, and newer GenAI features across its apps. ### Why focus on inference instead of training? Inference is the moment a trained model actually does work for users — ranking a Reel, picking an ad, answering a prompt, or generating text. That is different from training, which is the expensive one-time or periodic process of building the model in the first place. Meta’s bet is that custom chips pay off fastest on inference because those compound across billions of daily requests. ### Why are there four chips? Basically, Meta is saying one chip shape will not fit every AI job. The company described a portfolio approach — matching different silicon to different workloads — and the four-generation roadmap reflects that. Some MTIA generations are meant for ranking and recommendations, while later ones expand toward generative AI inference, where memory bandwidth and latency too. ### What makes this unusual? The speed. Meta said it plans four generations in two years, which is far quicker than the slower, monolithic chip cycles the industry is used to. That matters because AI workloads are changing faster than classic server hardware roadmaps can keep up with. Meta’s answer is iterative silicon — ship, learn, revise, repeat — instead of waiting years for one giant perfect design. ### Is Meta replacing Nvidia? No — and this is the important nuance. Meta is still buying lots of outside hardware and has explicitly said it uses a mix of Nvidia GPUs, AMD GPUs, CPUs, and its own MTIA chips. It also announced fresh partnerships around Broadcom and AWS Graviton in April, which reinforces the point: MTIA is the center of Meta’s custom plan, but not the whole stack. ### Why build custom chips at all? Cost, power, and fit. General-purpose GPUs are incredibly capable, but they are also expensive and not always the most efficient tool for every production inference task. Meta said it already deploys hundreds of thousands of MTIA chips for inference across organic content and ads. That scale means even modest efficiency gains can translate into very large savings in power, hardware spend, and data-center capacity. ### What does Broadcom have to do with it? Meta’s April announcement made clear that Broadcom is a long-term co-development partner for multiple generations of next-gen MTIA chips. So this is “in-house” in the sense that Meta defines the architecture and workload targets, but it is not doing every part alone. That is how a lot of modern custom silicon works — the hyperscaler owns the roadmap, while specialist partners help turn it into manufacturable hardware. ### Bottom line? Meta is trying to make AI serving look less like renting ever more giant GPU fleets and more like building a tailored hardware stack for its own apps. The four-chip MTIA roadmap matters because it turns that idea into an actual schedule — and a pretty aggressive one. If Meta can keep the cadence up, the bigger shift is not one flashy chip launch. It is inference becoming its own custom-silicon battleground.